Claude Mythos Finds New HAWK and Reduced-Round AES Attacks

Anthropic's Model Found Two Real Cryptographic Attacks. Neither Breaks Anything You Use.

Claude Mythos Preview halved the effective key size of a post-quantum signature scheme in 60 hours, then made an attack on reduced-round AES up to 800 times faster. Anthropic says no production software has to change.

Jul 29, 2026Source: Anthropic

Anthropic's Frontier Red Team published two pieces of novel cryptanalysis on 28 July, both produced by an unreleased model called Claude Mythos Preview. The results are real, checkable and narrow — and the company says so itself: "neither of these results has a practical impact on today's computer systems; no production software will have to change as a result."

HAWK

HAWK is a third-round NIST post-quantum signature candidate that had, in Anthropic's words, survived "two rounds of expert human review over a period of two years." In 60 hours of work the model found a nontrivial lattice automorphism that cuts the effective key size by a factor of two — meaning key sizes must double for the same security. Concretely, expected full key recovery against the small HAWK-256 parameter set drops from 2^64 to 2^38; gate-count estimates fall from 2^150 to 2^108 for HAWK-512 and from 2^288 to 2^182 for HAWK-1024.

AES, seven rounds of ten

The second result came from about three days of near-autonomous work and roughly one billion output tokens, with three substantive human prompts. The model devised a fingerprinting algorithm it named "Möbius Bridge" that removes the need to enumerate 2^56 values, making an existing meet-in-the-middle attack 200 to 800 times faster. It applies to a reduced 7-round research variant of AES-128, untouched since 2013, and still demands 2^105 chosen plaintexts. Full ten-round AES is unaffected.

The cost line

The number that makes this a business story rather than a maths one: each result cost roughly $100,000 in API spend to develop. Verification was the bottleneck — two researchers took nearly a month to gain confidence in the AES result. Anthropic shared the HAWK attack with its authors in June and disclosed to the public NIST mailing list at publication; a third, preliminary result recovers a 13-round LEA key in under an hour on a desktop, against a cipher deployed at 24 rounds.

Not yet reproduced

Validation so far is in-house. No independent reproduction has been published, and the NIST thread had no replies when outside reporters checked it.

Why It Matters

Cryptanalysis has always been rate-limited by the number of people able to do it. If a genuine new attack now costs about $100,000 of inference plus a month of human checking, the constraint moves from expertise to budget — and every unbroken scheme becomes a question of who is willing to spend.